SynthID Explained: How Google's AI Watermarking Works

SynthID Article Summary

1. SynthID embeds an imperceptible mark directly into content generated by compatible AI models.
2. The presence of the watermark provides information about the technical origin of the content, but it cannot determine whether the content is true, reliable, or misleading.
3. The absence of SynthID does not prove that content was created by a human, since it may have come from another generator or have been modified enough to make the watermark undetectable.

A perfectly natural-sounding voice calls you to confirm an appointment. A spectacular photograph circulates on LinkedIn. A video attributed to an executive appears just minutes before a sensitive announcement. In each case, the same question arises: how can you tell whether artificial intelligence played a role in creating the content?

Visible clues are no longer always enough. SynthID takes a different approach: embedding an invisible mark at the very moment AI generates text, an image, a video, or audio content. This trace can then be identified by a compatible detector.

What Is SynthID?

SynthID is a technology developed by Google DeepMind to embed a digital watermark into content produced or modified by artificial intelligence. The watermark is imperceptible to humans: it does not appear as a visible logo on an image, an audible sound in a recording, or a notice added to the bottom of a text[2].

The mark is incorporated into the structure of the content itself. A detection system configured to recognise it can then determine whether it is likely to be present.

SynthID therefore performs two related functions:

• watermarking, performed while the content is being generated;
• detection, performed when a user or system wants to examine the content's origin.

This approach differs from general-purpose AI detectors. Those systems analyse the characteristics of text, an image, or an audio file to estimate whether it resembles artificially generated content. SynthID instead looks for a signature intentionally embedded by the generator.

Why Did Google DeepMind Create SynthID?

The first experimental version of SynthID was introduced in August 2023 for images created with Imagen in Vertex AI. At the time, the system relied on two deep learning models trained together: one added the watermark, while the other searched for it in the image[3].

The technology was subsequently extended to text, audio, and video. It is used in particular with content generated by Gemini, Imagen, Lyria, and Veo. By May 2025, more than 10 billion pieces of content had already been watermarked, according to the figure announced when the SynthID Detector portal was unveiled[4].

The goal is not simply to identify malicious deepfakes. Watermarking can also help with:

• correctly attributing a creation;
• enforcing platform rules;
• moderating synthetic content;
• verifying a file received from a third party;
• providing transparency about the use of AI in an editorial process.

However, the system remains dependent on adoption. A generator that does not implement SynthID will not leave a SynthID watermark in the content it produces.

How Does SynthID Work?

The general principle is to introduce a signal while the content is being created. The form this signal takes varies depending on whether the content is an image, video, audio file, or text.

The watermark must meet two sometimes contradictory requirements: it needs to remain subtle enough not to degrade the output, while being robust enough to remain detectable after common transformations.

SynthID for Images and Videos

For images, SynthID subtly modifies certain information contained in the pixels. These variations are not intended to affect perceived image quality, but together they form a pattern that the detector can search for.

The same principle applies to video segments. The watermark is added during generation and designed to withstand common operations such as:

• cropping;
• adding filters;
• lossy compression;
• changing the frame rate;
• certain changes to colour or dimensions[2].

This resistance has practical value when an image is downloaded from a social network, compressed by a platform, or lightly edited before being reposted.

However, it should not be interpreted as a guarantee that the watermark will persist indefinitely. Numerous or substantial transformations can eventually weaken the signal to the point where it becomes undetectable.

SynthID for Audio Content

In an audio file, the watermark is embedded as an inaudible signal. It can be applied, for example, to content generated with the Lyria music model or NotebookLM's podcast generation feature[2].

The watermark is designed to remain recognisable after several common modifications, including:

• MP3 compression;
• adding noise;
• changing playback speed;
• certain processing or editing operations.

Detecting SynthID in a recording means that at least part of the file contains a signal originating from a compatible system. It does not verify the identity of a voice, the accuracy of what was said, or the context in which the recording was distributed.

SynthID for AI-Generated Text

Text contains neither pixels nor an acoustic signal. SynthID Text therefore operates when the model successively chooses the words—or, more precisely, the tokens—that make up its response.

At each step, the model assigns probabilities to the different tokens that could follow the text already generated. SynthID slightly adjusts these probabilities to create a statistical signature distributed throughout the response. It does not simply add a secret phrase or a sequence of hidden characters.

An experiment involving nearly 20 million responses with and without watermarks found no statistically significant difference in positive or negative user ratings. The method does not require changes to the model's training or access to the original model for subsequent detection, provided the appropriate watermarking configuration is available[5].

Detection is more effective on long and varied responses, such as an essay, a script, or several versions of an email. However, it becomes more difficult with a very short factual response because the model has fewer possible choices for embedding the signal without changing the meaning.

A complete rewrite or translation can also significantly reduce the detector's confidence level[6].

How Can You Detect Content Watermarked With SynthID?

The most accessible method is to use the Gemini app. A signed-in user can upload an image, video, or audio file and ask, for example: “Was this content created or modified by Google's AI?”

The process involves three steps:

1. Upload a single file to the conversation.
2. Ask an explicit question about whether it was created or modified by AI.
3. Review the result and, for certain media, the portions where the watermark was detected.

The file must currently be no larger than 100 MB. A video must be under 90 seconds long, and audio content must be under one hour. Approximate quotas also apply over a rolling 24-hour period[7].

This documented verification feature in Gemini applies to images, videos, and audio files. It should therefore not be presented as a public service capable of checking any text pasted into a conversation in the same way.

The SynthID Detector portal announced in 2025 was intended to bring detection for different formats together and highlight the portions most likely to contain the watermark. Its official page still mentions collaboration with journalists and media professionals to test the portal[2]. Its exact availability should therefore be checked before recommending the tool to a general audience.

How Should You Interpret the Result?

There are three possible outcomes.

A watermark is detected. All or part of the file was probably created or modified by an AI tool compatible with SynthID. In Gemini, verification currently identifies content generated by Google's AI tools, although other companies have begun adopting the technology[7].

No watermark is detected. The file was not recognised as content watermarked by the system being examined. However, it may still have been generated by another AI, produced by a tool that does not use watermarking, or subjected to transformations that weakened the watermark.

The result is uncertain. The content may be too simple, contain too little usable information, or have been modified only to a limited extent. Rather than forcing a binary answer, the system indicates that it does not have enough information.

In practice, a SynthID result primarily answers this question: “Does this file contain a technical mark originating from a compatible system?” On its own, it does not answer the question: “Is this content true?”

Is SynthID Reliable?

SynthID's reliability depends first and foremost on the content being examined and the transformations it has undergone.

A resized or compressed image can generally retain its watermark. The same applies to certain colour changes, video edits, or audio processing. However, after a succession of transformations, the watermark may no longer be detectable[7].

For text, length and freedom of expression matter significantly. A detailed response provides more opportunities to embed a statistical signature than a city name, a required quotation, or a very short sentence. The signal can also be weakened by:

• extensive paraphrasing;
• translation;
• rewriting by another model;
• replacing a large portion of the vocabulary.

The scientific publication on SynthID Text itself emphasises that generative watermarks are not a complete solution. They require coordination between model providers and remain vulnerable to attempts to remove, imitate, or forge the signal[5].

Two interpretation errors should therefore be avoided:

• treating the absence of a watermark as proof of human creation;
• treating its presence as proof that the information being presented is accurate.

An authentic photograph can be misleading if published with a false caption. Conversely, an AI-generated illustration can legitimately accompany content when it is clearly identified as such. SynthID provides information about how content was produced, not about whether its message is true.

SynthID, AI Detectors, and Content Credentials: What's the Difference?

These technologies share a common goal of improving transparency, but they do not provide the same information.

Content Credentials are based on the open C2PA standard. They function as a provenance history cryptographically linked to the file and can record its origin, transformations, or use of AI. However, they do not certify that the content itself is true. Their main purpose is to verify that provenance information is valid, signed, and has not been altered[8].

SynthID and Content Credentials can therefore complement one another. The former looks for a signal embedded in the content; the latter documents its history.

Can SynthID Detect a Phone Call Made by AI?

SynthID can help identify an artificial voice in a call recording, but only if that voice was generated by a tool that applies the watermark. At this stage, it cannot universally detect all calls made by AI agents in real time.

This distinction is essential. SynthID analyses the provenance of an audio signal. On its own, it does not determine who or what is controlling the conversation.

For example, a call may be handled by an AI agent using a synthetic voice without a watermark. In that case, SynthID will find nothing. Conversely, a real person could play a prerecorded message watermarked with SynthID during a call. The presence of the watermark would indicate synthetic audio content, but not necessarily a fully automated call.

When Is Detection Possible?

Three conditions must be met:

1. The voice must come from an engine that applies SynthID.
2. The watermark must remain sufficiently intact after telephone transmission.
3. The recording must be analysed with a detector compatible with the configuration used by the generator.

The scope is beginning to extend beyond Google's own tools. In June 2026, ElevenLabs announced the integration of SynthID into its Text to Speech generations, with a gradual expansion to its other audio content. At the same time, the company offers a free detector designed to verify whether a clip was generated by its own tools[9].

This development makes it conceivable to analyse a call using an ElevenLabs voice. However, it does not guarantee that all calls originating from its voice agents can be identified under all conditions. Available documentation does not confirm the watermark's resistance to every codec, transcoding process, filter, source of noise, and processing operation applied by telephone networks.

Can AI Be Detected During the Call?

The documented public tools currently work from an uploaded audio file. Gemini allows users to submit a recording and then ask whether it was created or modified by Google's AI. It does not recognise content from every provider and is not presented as a live call-monitoring system[7].

In a contact centre, a realistic use of SynthID would therefore look more like this:

• record the call in compliance with applicable rules;
• isolate a section containing the voice suspected of being artificial;
• analyse it using the suspected provider's detector;
• cross-check the result against the caller's number, conversational scenario, metadata, and information disclosed by the caller.

A positive result can attribute the voice to a compatible system. A negative result does not prove that a human was on the other end of the line.

SynthID Is Not a Universal Voice-Agent Detector

To automatically identify AI calls, a company would need a broader system combining several signals:

• audio watermark detection;
• acoustic analysis;
• detection of automated conversational behaviour;
• caller authentication;
• information provided by the carrier or service provider;
• explicit disclosure of AI use.

The last of these remains the clearest solution for the person on the other end of the call. European transparency rules also provide that people should be informed when they are interacting directly with certain AI systems[1]. In a business call, stating at the outset that the call is being handled by an AI voice agent is more reliable and understandable than expecting the recipient to conduct their own technical investigation.

How Can Businesses Use SynthID?

For businesses, SynthID's main value lies in verifying content whose production chain is not entirely under their control.

Checking an Image Provided by a Vendor

A marketing team receives a campaign containing several images. The company can request the original files and their provenance information, check for a potential watermark, and then document the tool used and the associated rights.

The SynthID result does not replace license verification or checks concerning the people, brands, and works depicted. It adds a technical indicator to the validation process.

Reviewing a Video Before Publication

A short video attributed to an executive is circulating on social media. A check in Gemini can search for a watermark from Google's tools. If no watermark is found, the team should continue its investigation by examining the source of the first upload, the account that published the file, official communications from the person concerned, and any possible editing cuts.

The key micro-insight is simple: a negative result should start the investigation, not end it.

Verifying Audio Content

A team receives a voice clip intended for use in training materials or an advertisement. Detection can identify audio content produced by a compatible tool. The team must then verify authorisation to use the voice, the consent of the individuals involved, and the accuracy of the message.

Establishing Rules for Editorial Production

For text, SynthID should not become a disciplinary tool used in isolation against an employee, candidate, or student. Short, edited, or translated texts provide less favourable detection conditions, while incompatible generators remain outside the system's scope.

A more robust internal policy can specify:

• permitted uses of AI;
• content that requires human review;
• provenance information that must be retained;
• disclosures intended for the public;
• the procedure to follow when content appears questionable.

This governance complements the practices described in our guide to using artificial intelligence in business.

Key Takeaways About SynthID

SynthID provides a precise technical response to a clearly defined problem: identifying within a piece of content the mark added by a compatible generator. Its main advantage comes from embedding that information at the source, which is more informative than simply estimating whether a text or image looks AI-generated.

That precision also means its limitations must be respected. A positive detection provides information about technical provenance, not truthfulness. A negative detection does not rule out the use of another AI system. For businesses, the most reliable approach is therefore to combine watermarking, Content Credentials, contextual analysis, retention of original files, and human review.

To explore the subject further, see our resources on artificial intelligence and AI agents.

SynthID FAQ

Can SynthID Detect All AI-Generated Content?

No. SynthID detects content to which a compatible system has previously added its watermark. Content from an incompatible model can be entirely AI-generated without carrying this mark.

How Can You Check an Image With SynthID?

Upload the image to Gemini using an eligible account, then ask whether it was created or modified by Google's AI. Gemini searches for the watermark and may indicate the areas where it is most likely to be present[7].

Can a SynthID Watermark Be Removed?

Common modifications are not necessarily enough to remove it. However, a succession of substantial transformations can reduce its detectability. Text is particularly vulnerable to extensive rewriting and translation[6].

Is the SynthID Watermark Visible?

No. It is designed to be imperceptible in images and videos, inaudible in audio files, and invisible in the wording of text. Only a compatible detector can search for the signal.

Is SynthID Enough to Comply With the AI Act?

SynthID can contribute to the machine-readable marking expected of certain providers of AI systems. However, it does not constitute a certification of compliance.

Article 50 distinguishes between several responsibilities. Providers must, among other things, enable the detection of certain synthetic content, while deployers must inform people exposed to certain deepfakes or artificially generated text on matters of public interest[1].

A company must therefore consider its role, the nature of the content, and the context in which it is distributed. An invisible watermark does not automatically eliminate the need for a visible disclosure or information that the public can understand.

What Is the Difference Between SynthID and a Visible Watermark?

A visible watermark immediately informs the reader or viewer, but it can often be cropped or removed. SynthID is embedded within the content and is designed to withstand certain modifications, but detecting its presence requires a compatible tool.

Does SynthID Prove That Content Is Fake?

No. The watermark provides information about the use of generation or modification technology. It does not verify the facts presented, the context, or the author's intent.

Can SynthID Be Used to Detect a Call Made by AI?

SynthID can detect a watermark in a call recording if the voice was produced by a compatible generator. However, it does not detect every voice agent and does not function as a universal real-time identifier. The absence of a watermark therefore does not allow you to conclude that the caller was human.

What Is C2PA?

C2PA is a coalition behind an open standard for tracing the provenance of digital content. It is based on Content Credentials, which record the origin of a file, the tools used, and certain modifications in a cryptographically verifiable way. Unlike SynthID, C2PA does not search for an invisible watermark: it documents the content's history without guaranteeing that the information it presents is true[8].

Citations

• [1]https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations
• [2]https://deepmind.google/models/synthid/
• [3]https://deepmind.google/blog/identifying-ai-generated-images-with-synthid/
• [4]https://blog.google/innovation-and-ai/products/google-synthid-ai-content-detector/
• [5]https://www.nature.com/articles/s41586-024-08025-4
• [6]https://deepmind.google/blog/watermarking-ai-generated-text-and-video-with-synthid/
• [7]https://support.google.com/gemini/answer/16722517?hl=fr
• [8]https://c2pa.org/specifications/specifications/2.2/explainer/Explainer.html
• [9]https://www.theverge.com/ai-artificial-intelligence/957510/elevenlabs-rolls-out-synthid-support

Published on October 5, 2026.